Skip to content
kode-gopher

Getting started

This section runs kode-gopher on your machine. Your MCP client (Claude Code, Claude Desktop, Gemini CLI) starts kode-gopher serve as a subprocess over stdio. Each snippet the model writes is built and run in a sandbox pod on a Kubernetes cluster that your kubectl can reach. Snippets use your own Google credentials.

To run kode-gopher as a shared server for a team instead, see Deploy for a team.

What you need

Section titled “What you need”
  • Go 1.26 or later, to install the CLI. Go downloads the exact toolchain version it needs.
  • A Kubernetes cluster with agent-sandbox: a local kind cluster (Docker required), or GKE Autopilot with the agent-sandbox addon. Set up a sandbox cluster covers both.
  • kubectl, with a context for that cluster and permission to create, patch and delete SandboxClaims (the patch renews each claim’s lease) and to port-forward in its namespace.
  • gcloud, signed in to Application Default Credentials (ADC):
    Terminal window
    gcloud auth application-default login
    export GOOGLE_CLOUD_PROJECT=<your-project>

Steps

Section titled “Steps”
  1. Install the CLI.
  2. Set up a sandbox cluster, and check it from the command line.
  3. Connect your MCP client, and ask the model to use Google Cloud.